bench.txt

PASS
BenchmarkShift-4	   20000	     69666 ns/op	 940.71 MB/s
BenchmarkAnd-4  	   20000	     66802 ns/op	 981.04 MB/s
ok  	_/c_/Temp/shifttest	4.146s
PASS
BenchmarkShift-4	   20000	     68680 ns/op	 954.21 MB/s
BenchmarkAnd-4  	   20000	     67339 ns/op	 973.21 MB/s
ok  	_/c_/Temp/shifttest	4.121s
PASS
BenchmarkShift-4	   20000	     68687 ns/op	 954.12 MB/s
BenchmarkAnd-4  	   20000	     67786 ns/op	 966.80 MB/s
ok  	_/c_/Temp/shifttest	4.217s

shift_amd64.go

package shifttest

func reference(a, b []uint64) {
	for i := range a {
		a[i] = (a[i] & 0x1ff) | (b[i] & 0xfffffffffffffc00)
	}
}

func testShift(a, b []uint64)

func testAnd(a, b []uint64)

shift_amd64.s

// func testShift(a, b []uint64)
TEXT ·testShift(SB), 7, $0
	MOVQ a+0(FP), SI      // SI: &a
	MOVQ a_len+8(FP), R10 // R10: len(out)
	MOVQ b+24(FP), DX     // DX: &b

shloop:
	MOVQ (SI), AX
	MOVQ (DX), BX

	SHLQ $10, AX  // a <<= 10
	SHRQ $54, BX  // b >>= 54
	SHRQ $10, AX  // a >>= 10
	SHLQ $54, BX  // b <<= 54
	ORQ  AX, BX   // b |= a
	MOVQ BX, (SI) // *a = b

	ADDQ $8, SI // move to next
	ADDQ $8, DX // move to next
	DECQ R10
	JNZ  shloop
	RET

// func testAnd(a, b []uint64)
TEXT ·testAnd(SB), 7, $0
	MOVQ a+0(FP), SI      // SI: &a
	MOVQ a_len+8(FP), R10 // R10: len(out)
	MOVQ b+24(FP), DX     // DX: &b

andloop:
	MOVQ (SI), AX
	MOVQ (DX), BX

	MOVQ $0x1ff, R9
	ANDQ R9, AX
	MOVQ $0xfffffffffffffc00, CX
	ANDQ CX, BX
	ORQ  AX, BX                  // b |= a
	MOVQ BX, (SI)                // *a = b

	ADDQ $8, SI  // move to next
	ADDQ $8, DX  // move to next
	DECQ R10
	JNZ  andloop
	RET

shift_test.go

package shifttest

import (
	"testing"
)

/*
func TestConstants(t *testing.T) {
	fmt.Printf("%x\n", uint64(((1<<64)-1)-((1<<10)-1)))
}
*/

func BenchmarkShift(b *testing.B) {
	const size = 64 << 10
	var x [size]uint64
	var y [size]uint64
	b.SetBytes(size)
	// Warm up
	testShift(x[:], y[:])
	b.ResetTimer()
	for i := 0; i < b.N; i++ {
		testShift(x[:], y[:])
	}
}

func BenchmarkAnd(b *testing.B) {
	const size = 64 << 10
	var x [size]uint64
	var y [size]uint64
	// Warm up
	testAnd(x[:], y[:])
	b.SetBytes(size)
	b.ResetTimer()
	for i := 0; i < b.N; i++ {
		testAnd(x[:], y[:])
	}
}